Football Video Semantic Analysis

Content based multimedia retrieval has been a hot research field recent years. As a part, sports video retrieval has been studied widely. Soccer is a very popular game, which has a large number of audiences all over the world. Therefore, this paper used soccer video from TV program as an example, extracting high-level semantics, such as the event when goal shot occurs. This paper presents a framework for soccer video semantic analysis, consisted of the parts as follows: video segmentation based on cut detection; field segmentation; shot classification; slow-motion shots detection; forbidden zone detection; caption region extraction and exciting events detection.Video segmentation is the first step, this paper presents an algorithm to segment video by cut shot detecting, for most of shots transitions in soccer video are cut. Select key frames in video segments, and detect field region in each frame. Shots are classified into four types: long-view shot, middle-view shot, close-up shot and out of field shot, using field rate, shape and areas of player in the field. Slow-motion replays often appears after shoot shoting or foul occurs, so we determine if a given shot consists of a slow-motion replay to detect events. This paper presents an algorithm for slow-motion replays detecting. As shoot often happens near forbidden zone, and forbidden zone lines in a long-view shot can be detected to find forbidden zone, which can help shoot event detection. If a goal occurs, the score will change, so there will be score caption in the video. Therefore, score captions are used to help detect shoot event and determine whether goal happens.In a word, this paper presents an algorithm triggered by slow-motion replays. With the results, go back certain seconds to find temporal sequences of shots to detect exciting events, assisted by forbidden zone detection and caption region extraction. This paper implements a prototype system for automatic soccer video semantic analysis by Visual C++6.0. Experiments have demonstrate that the algorithm is effective.